How I Built My Own AI Agents to Run Parts of My Business

It started with a problem, not a plan

A few weeks ago I launched AI Work Academy as a side project while keeping my full-time job. I quickly realized that building the course was only half the work. The other half was everything around it: checking analytics, figuring out what to post on social media, keeping up with AI news so the content stays relevant, thinking about SEO. All of that on top of a regular workday.

I'm one person. I don't have a marketing team, a social media manager, or a newsletter editor. But I thought: what if I could build AI agents to fill those roles? Not as a gimmick, but as actual help. Like hiring someone who has more expertise than you in a specific field, except the "someone" is an AI that runs on a schedule and costs a few cents per run.

That reframing changed everything. I wasn't just automating tasks. I was giving myself a Head of Marketing that analyzes real data and drafts strategy, and a News Curator that reads more sources in ten minutes than I could in a week. That's when I decided to actually build it.

What an AI agent actually is (no buzzwords)

Before I get into what I built, let me clear up what I mean by "AI agent" because the term gets thrown around a lot. An AI agent is not a chatbot. It's not a prompt you run manually. It's a program that runs on its own, on a schedule, and uses AI to make decisions about what to do.

Think of it like this: when you ask ChatGPT to rewrite an email, that's you using AI. When a program wakes up at 10am every two days, reads the latest AI news from a list of trusted sources, writes a newsletter about it, and sends you a Telegram message with a link to read it, that's an agent.

The key difference is autonomy. You set it up once, give it clear instructions about what it should do and what tools it can use, and then it runs without you sitting at the keyboard.

My setup: three agents, one platform

I built a small platform I call AgentHub. It runs on a cloud server and has a dashboard where I can see what each agent is doing, read their output, and trigger runs manually when I need something specific.

Right now I have three agents:

The first is an AI News Curator. Every two days it sweeps a list of trusted sources: official blogs from Anthropic, OpenAI, DeepMind, Meta, and others, plus research papers and a small list of reliable accounts on X. It writes a newsletter covering what happened in the AI world over the last seven days. No opinions, no hype, just factual reporting with links to primary sources. It sends me a short summary on Telegram and saves the full version as a document I can read from the dashboard.

The second is a Head of Marketing agent. This one is more complex. Every two days it pulls real data from Vercel Analytics and Google Search Console using API integrations, reviews how recent content performed, updates the marketing strategy, and queues draft posts and engagement actions for the next 48 hours. It also writes long-form strategic documents when something bigger is needed, like a video brief or a campaign plan.

The third is a Social Media Manager that was originally meant to execute the marketing plans, but I ended up disabling it. More on why in a moment.

The part nobody talks about: why agents can't do everything

Here's something you won't see in most AI content: my agents can't post to social media. Not because of a technical limitation in the AI itself, but because platforms like X and Instagram are very good at detecting automated logins from cloud servers. I tried using browser automation to log in and post from the server. X blocked it within days.

So I had to rethink the architecture. Instead of trying to make the agent do everything end to end, I split the work: the agent plans, I execute.

The Head of Marketing drafts content and queues it. I review the queue from my phone via Telegram, approve or reject items, and when I'm ready, I run a single command on my laptop that picks up the approved items and posts them through my real, logged-in browser. The agent handles the thinking. I handle the last mile.

This turned out to be a much better design anyway. I stay in control of what gets published, the agent handles the research and strategy work that was eating my mornings, and nothing gets posted that I haven't reviewed.

How the news agent changed my relationship with AI news

Before I built the AI News Curator, I was doing what most people in tech do: scrolling X for an hour, clicking through dozens of tabs, trying to figure out what was real news and what was hype. It was exhausting, and I still felt like I was missing things.

Now, every two days, I get a clean summary on Telegram. Five to ten minutes of reading instead of an hour of scattered browsing. The agent only covers the last seven days, so nothing stale sneaks in. It attributes every claim to its source, so I can click through to verify anything that seems important. And it separates actual product launches and research papers from speculation and commentary.

It's the kind of thing that sounds simple but has genuinely improved my day. I'm better informed now than when I was spending three times as long trying to keep up manually.

And this got me thinking: if this is useful for me, it might be useful for other people too.

What I actually learned building this

Building AgentHub taught me things I couldn't have learned from a tutorial. The biggest lesson was about trust boundaries. When you give an AI agent access to tools (databases, APIs, a browser), you have to think carefully about what it should and shouldn't be able to do. I ended up building an environment isolation system where each agent gets its own set of encrypted credentials and can't see another agent's secrets.

The second lesson was about memory. An agent that runs every two days needs to remember what it did last time. Otherwise it drafts the same content twice, repeats the same analysis, or misses context about an ongoing campaign. I built a memory system where agents save learnings and context after each run, and those memories get loaded into the next run's prompt.

The third lesson was about scheduling. Not every task needs to run every day. My agents used to run daily, and I was drowning in output I couldn't keep up with. Switching to every two days gave each run a real 48-hour window of data to work with and stopped the queue from piling up faster than I could review it.

None of this is rocket science. But you only learn it by building something real and watching it run in production.

You don't need to build what I built

I want to be clear: you don't need to build your own agent platform to benefit from AI agents. I built AgentHub because I work with technology and I enjoy figuring out how to make tools work for me. But the principles are the same whether you're using Claude's built-in features, Make.com workflows, or any other automation platform.

The real skill is learning to break your work into pieces and figuring out which pieces an AI can handle autonomously. That's what Module 5 of AI at Work Academy covers: building reusable skills and scheduling them. And Module 9 goes all the way to creating your own AI agent.

The point of sharing my setup isn't to impress you. It's to show you what's possible when you actually apply the things we teach in the course. These aren't theoretical concepts. They're tools I use every day to run a real business.

Would you read an AI news brief written by an agent?

Here's the thing I keep coming back to: the AI News Curator writes genuinely useful summaries. Factual, well-sourced, no fluff. I read every one of them. And several people I've shared them with have told me they'd subscribe to something like that.

So I'm testing the idea. If enough people are interested, I'll start sending out the AI news brief as a newsletter: a curated summary of what's happening in AI, delivered to your inbox every two days. Same quality I get from my own agent, same no-hype approach.

No commitment, no cost. I just want to see if there's real interest before I set it up. If you'd like to receive it, drop your name and email below and I'll reach out when it's ready.